Zeroes of the Partial Sums of the Exponential

Function
Ian Zemke
June 3, 2009

Contents
1 Introduction 2

2 Asymptotic Behavior of Zeroes and the Zero Free Region 2

2.1 Lower Bound on Distance to Γ . . . . . . . . . . . . . . . . . . . 3
2.2 Upper Bound on Distance to Γ. . . . . . . . . . . . . . . . . . . . 5

3 Characterization of the Exponential 8

3.1 Factoring an Entire Function . . . . . . . . . . . . . . . . . . . . 8
3.2 Characterizing the Exponential Function . . . . . . . . . . . . . . 11

4 Conclusion 12

1
1 Introduction
In this paper we present a summary of some surprising results about the zeroes
of partial sums of the exponential function. We define the function
n
X zk
sn (z) =
k!
k=0

to be the nth partial sum of the exponential function on the complex plane.
The Fundamental Theorem of Algebra shows that each sn has n zeroes count-
ing multiplicities, whereas the actual exponential function has no zeroes on the
entire complex plane. The paper falls into two parts. In the first part, we com-
pute directly the behavior of the zeroes of these partial sums. Since most of
the results from the first half are relatively unknown, we will go through the
proofs in detail. In the second half of the paper, we show how this behavior in
some sense uniquely characterizes the exponential function. Naturally this re-
quires considerably more sophistication, and in proving the certain statements,
we will need to introduce several large-caliber results including the Weierstrass
and Hadamard Factorization Theorems. All figures were produced using Math-
ematica.

2 Asymptotic Behavior of Zeroes and the Zero

Free Region
It becomes convenient to make the substitution

pn (z) = sn (nz).

We show in this section that the zeroes of pn asymptotically fall along a horse-
shoe shaped curve in the unit circle and we obtain estimates on the rates of
convergence. In this section we present a description of how the zeroes behave
asymptotically. Specifically, we show that the zeroes of pn (z) (the scaled partial
sums) asymptotically fall near the curve

Γ = {z : |ze1−z | = 1, |z| ≤ 1}.

A plot of Γ is shown in Figure 1 below.

2
 1.0

0.5

-1.0 -0.5 0.5 1.0

-0.5

-1.0

Figure 1: The cureve Γ.

Further we show that the zeroes always fall outside of the region bounded
by Γ, and we obtain estimates on how quickly the zeroes approach the curve.

2.1 Lower Bound on Distance to Γ

The following lemma and theorem are modified from [5, Theorem 1], which is a
refinement of an original result by J.D. Buckholtz in [1].
Lemma 2.1. For all n, pn (1) > en /2.
Proof. The proof is mainly a little bit of cleverness in evaluating power series.
Observe that
Z Z
1 x n −t 1 ∞ n −t
sn (x)e−x = t e dt = 1 − t e dt,
n! 0 n! x
which can be verified simply by evaluating the integrals using integration by
parts. Since we will be setting x = n, both integrals in the above expression
will be positive, and hence it is sufficient to show that
Z n Z ∞
tn e−t dtx < tn e−t dt. (1)
0 n

Through a change of variables, we see this is equivalent to

Z 1 Z ∞
−t n
(te ) dt < (te−t )n dt.
0 1

3
By examining Taylor coefficients and performing some algebraic manipulation,
we see that
eu (1 − u) < e−u (1 + u),
for all u such that 0 < u < 1. Hence if g(t) = te−t then we have

g(1 − u) < g(1 + u)

which implies that Z Z

1 2
(te−t )n dt < (te−t )n dt,
0 1

which clearly implies (1).

Theorem 2.2. There are no zeroes z of pn with |z| ≤ 1 and |ze1−z | ≤ 21/n .
Proof. Observe that it is sufficient to show that if |z| ≤ 1 and |ze1−z | ≤ 21/n ,
then |1 − e−nz pn (z)| < 1, which implies that pn (z) 6= 0. Compute as follows:
¯ ¯
¯ X∞
(nz) k¯
¯ ¯
|1 − e−nz pn (z)| = ¯e−nz ¯
¯ k! ¯
k=n+1
¯ ¯
¯¡ ∞ ¯
¯ 1−z ¢n −n X nk z k−n ¯
= ¯ ze e ¯
¯ k! ¯
k=n+1
∞
X nk
≤ 2e−n <1
k!
k=n+1

by the lemma, which thus implies that pn 6= 0.

Corollary 2.3. There is a number C such that sn has no zeroes in the circle
of radius Cn.

Theorem 2.4. There are no zeros z of pn for which dist(z, Γ) ≤ (21/n −1)/(2e2 )
and |z| ≤ 1.

Proof. Let f (z) = ze1−z . Hence f 0 (z) = (1 − z)e1−z . Then if dist(z, Γ) ≤ d =

(21/n − 1)/(2e2 ) and u is a point on Γ such that |z − u| ≤ d, then we have

f (z) ≤ |f (u)| + 2e2 d = 21/n

so that applying the result from the previous theorem shows that there are no
zeroes within 21/n of Γ which lie inside the unit disk.
This theorem sets a bound on how quickly the zeroes of pn can approach
the curve Γ. At this point, we still haven’t shown that the zeroes of sn even
approach Γ, but we have shown that the region bounded by Γ is free of zeroes
and that the zeroes approach no faster than 21/n .

4
2.2 Upper Bound on Distance to Γ.
The next natural question to ask is how quickly the zeroes of pn approach Γ.
We now present a several lemmas leading to a theorem due to Buckholtz in [1]
and [2]. Let Tn be defined as

n!
Tn (z) = pn (z).
(nz)n

Lemma 2.5. The following equation is true for all z 6= 0, 1.

µ ¶
z T 0 (z)
Tn (z) = 1+ n , z 6= 0, 1. (2)
z−1 n

Proof. It is sufficient to show that

µ ¶
Tn0 (z)
(z − 1)Tn (z) = z 1 + .
n

Computing the left side of this equation yields

(z − 1)Tn (z) = zTn (z) − Tn (z)

Xn n
nk−n n! k−n+1 X nk−n n! k−n
= z − z
k! k!
k=0 k=0
∞
X k−n−1
n n!
=z− (k − n) z k−n .
k!
k=0

Similarly, evaluating the right side of the equation yields that

µ ¶ Ã " n #0 !
Tn0 (z) 1 X nk−n n! k−n
z 1+ =z 1+ z
n n k!
k=0
n
X nk−n−1 n!
=z+ z k−n ,
k!
k=0

which is clearly equal to what we get on the right hand side. Therefore (2)
holds.

Lemma 2.6. If |ze1−z | ≥ 1, then

√
|Tn (z)| < 2e n.

Proof. The proof is relatively straightforward, but involves defining the function
Sn as
X∞
n! (nz)k
Sn (z) = .
(nz)n k!
k=n+1

5
 A straightforward manipulation yields that yields that

n!(e/n)n
Tn (z) + Sn (z) = . (3)
(ze1−z )n

Applying the triangle inequality to the functions Tn and Sn yields that for
|z| ≤ 1, we have
|Sn (z)| ≤ Sn (1)
and similarly for |z| ≥ 1, we have

|Tn (z)| ≤ Tn (1).

Further, since Tn (1) and Sn (1) are both positive both must be smaller than
√
n!(e/n)n ≤ e n,

where the latter inequality follows from Sterling’s inequality. Applying the
we see if |ze1−z | ≥
triangle inequality to the cases |z| ≤ 1 and |z| ≥ 1 separately,√
1, then both of Tn and Sn are smaller in magnitude than 2e n.

√
Theorem 2.7. For every integer n, the zeroes of of pn lie within distance 2e/ n
of Γ.

Proof. Notice that the zeroes of pn are the same as those for Tn .
Since Tn is analytic outside of Γ and the region bounded by Γ is convex,
we can use the Cauchy’s Inequality for derivatives with the above lemma to see
that if z is δ or farther away from Γ (and outside of the region bounded by Γ),
then we have √
|Tn0 (z)| ≤ 2e n/δ.
√
Inserting this into (2) yields if δ > 2e/ n, we have no zeroes for Tn and thus
none for pn .
Together, the preceding theorems give us a good understanding of the behav-
ior of the zeros of pn . They show that the zeroes asymptotically and uniformly
approach Γ and give us estimates on the rates of convergence. Below we show
several renderings of the behavior of the zeroes of pn for various values of n.

6
 0.6
0.6

0.4 0.4

0.2 0.2

-0.4 -0.2 0.2 0.4 0.6 0.8 1.0 -0.4 -0.2 0.2 0.4 0.6 0.8 1.0

-0.2 -0.2

-0.4 -0.4

-0.6
-0.6

(a) Zeroes for n = 3, shown next to Γ. (b) Zeroes for n = 5, shown next to Γ.

0.4

0.2

-0.2 0.2 0.4 0.6 0.8 1.0

-0.2

-0.4

(c) Zeroes for n = 3, shown next to Γ.

Figure 2: Zeroes of pn for n = 3, 5, 10.

7
 1.0

0.5

-1.0 -0.5 0.5 1.0

-0.5

-1.0

Figure 3: The n = 10 case with the proven outer and inner boundaries for the
zeroes shown. The boundaries are rendered as dashed lines. Γ is shown as a
solid line.

3 Characterization of the Exponential

A surprising feature of the exponential is that just a small amount of informa-
tion on the behavior of the zeroes uniquely determines the exponential function.
It is not surprising that knowing that the zeroes of pn asymptotically fall along
Γ would uniquely determine the exponential function, but perhaps more strik-
ing is that we can extend Corollary 2.3 to be an if and only if statement. In
the following subsection, we present a discussion leading to the statements of
the Weierstrass and Hadamard factorization theorems. We will discuss these
theorems and give proofs where simple, but more involved proofs of more well
known results are not of particular interest for this paper and thus can be found
in [3] by the interested reader. These two theorems will lead to a result proved
by Buckholtz in [1].

3.1 Factoring an Entire Function

For our discussion of the zeroes of the partial sums of the exponential function,
we wish to be able to think of factoring an entire function which may have an
infinite number of roots. The simplest of all cases is that of an entire function
with no zeroes.

8
Theorem 3.1. Let f (z) be an entire function with no zeroes. Then f (z) = eg(z)
for some entire function g .

Proof. Let Z z
f 0 (ζ)
g(z) = dζ + Logf (0),
0 f (ζ)
which we can observe the be analytic since the f (z) is never zero, so the in-
tegrand is analytic. It is easy to verify that eg(z) and f (z) both solve the
differential equation
f 0 (z)
h0 (z) = h(z)
f (z)
with h(0) = f (0). Therefore by the uniqueness principle for differential equa-
tions, f and eg must be the same function.
To consider functions that actually have zeroes, we need to introduce some
machinery. Let

E0 (z) = (1 − z)
µ ¶
z2 zp
Ep (z) = (1 − z) exp z + + ··· + , p ≥ 1.
2 p

Theorem 3.2. If {an } is a sequence in C such that lim |an | = ∞ and an 6= 0

for n ≥ 1. If {pn } is any sequence of integers such that

X∞ µ ¶pn +1
r
<∞
n=1
|an |

for all r > 0, then

∞
Y µ ¶
z
P (z) = Epn
n=1
an
converges and is analytic. Furthermore, the sequence pn = n − 1 will always
yield convergence.

We usually call
∞
Y µ ¶
z
P (z) = Epn
n=1
an
the canonical product. The proof of this theorem is not particularly difficult,
and relies mostly on relatively simple estimates on the convergence of the prod-
uct, but is somewhat involved and will be omitted. The interested reader can
find it in [3] or many other elementary texts on complex analysis.
Applying the previous theorems yields the following formulation of the Weier-
strass Factorization Theorem.

9
Theorem 3.3 (The Weierstrass Factorization Theorem). Let f be an entire
function and let {an } be the zeroes of f repeated with appropriate multiplicity
and suppose f has a zero of order m at z = 0. Then there is a sequence of
integers pn and an entire function g such that
Y∞ µ ¶
z
f (z) = z m eg(z) Epn . (4)
n=1
an

Proof. The proof follows almost immediately from the previous theorems. From
Theorem 3.2 we know there’s a function h in the following form
Y∞ µ ¶
z
h(z) = z m Epn ,
n=1
an

which has the same zeroes as f with the same multiplicities. Therefore f /h is
entire and nonzero, so by Theorem 3.1, we know that
f (z)
= eg(z)
g(z)
for some entire function g(z), as we wanted to show.
We now present several definitions and basic theorems leading to a statement
of the Hadamard Factorization Theorem.
Definition. Let f be an entire function with zeroes {a1 , a2 , . . . } repeated ac-
cording to multiplicity and arranged such that |a1 | ≤ |a2 | ≤ · · · . Then we say
f is of finite rank if
X∞
|an |−(p+1) < ∞
n=1
for some integer p. We say f is of rank p if p is the smallest integer which
yields convergence in the preceding series.
Observe that if f is of finite rank then
∞
Y µ ¶
z
P (z) = Ep
n=1
an

will work as an expression for the canonical product in (4). Notice also that if
we use the above expression for P (z), then our expression for f is unique except
for integer multiples of 2πi in g(z). We say such an expression for P (z) is in
standard form.
Definition. An entire function f is said to have finite genus if f has finite
rank and if
f (z) = z m eg(z) P (z)
where P (z) is in standard form and g is a polynomial. If p is the rank of f and
q is the degree of g, then we call µ = max{p, q} the genus of f .

10
Definition. We say a function is of finite order if there is a constant α ∈ R
such that λ
|f (z)| ≤ e|z|
whenever |z| is sufficiently large. We call the infimum of all λ such that this is
true the order of f .
We now cite an important theorem relating the genus of a function to it’s
order, as is stated in [3].
Theorem 3.4. If f is an entire function of finite genus µ, then f is of finite
order and λ ≤ µ + 1.
The proof of this theorem involves several pages of computation, and thus
will be omitted. It should be noted though, that despite it’s length, the proof
is relatively elementary and involves only relatively straightforward estimation.
The interested reader can find it in [3].
We are finally ready to present the Hadamard Factorization as a converse to
the previous result. We need this factorization theorem to prove a characteri-
zation of the exponential function through the roots of the partial sums.
Theorem 3.5. Hadamard Factorization Theorem If f is an entire function of
finite order λ then f has finite genus µ ≤ λ.
Again, the proof of this theorem is relatively straightforward, but beyond
the interest of this paper. The interested reader can find it in [3].

3.2 Characterizing the Exponential Function

In this section we prove a theorem about how the growth of the zeroes of the
partial sums of the exponential function completely characterize the exponential
function. First we need to state a lemma on calculating the order of a function.
P∞
Lemma 3.6. If f (z) = n=0 an z n then the order of f can be calculated by

n log n
λ = lim sup .
n→∞ ln(1/|an |)

We won’t prove this lemma since the result is fairly well known. The inter-
ested reader can find a proof in [4] or the outline of the proof in an exercise in
[3]. The following theorem represents the culmination of this entire paper. It
originally was due to Buckholtz and can be found in [1].
P∞
Theorem 3.7. Suppose f (z) = k=0 ak z k is an entire function. The following
two statements are equivalent:
Pn
(i) There is a positive number c such that for each n, the function k=0 ak z k
has no zeroes with norm less than cn.
(ii) The function f can be represented as aebz .

11
Proof. We have already proved that (ii) implies (i), so we must just show the
other direction. Let f be an entire function such that
∞
X
f (z) = an z n
k=0

and let
n
X
sn (z) = ak z k .
k=0

Now suppose that there is a constant c such that sn has no zeroes in the disk
|z| < cn. By Hurwitz’s theorem, f has no zeroes. Since |an /a0 |1/n is the
geometric mean of the modulii of the zeroes of sn , and since all of the zeroes
are outside of the circle of radius cn, we know that
¯ ¯1/n
¯ an ¯
¯ ¯ > cn.
¯ a0 ¯

Taking logarithms and rearranging yields

log n + log c
1 .
n log |an | − n1 log |a0 |

Separating out terms that go to zero yields that

n log n
lim sup ≤ 1.
n→∞ log(1/|an |)

Therefore, by the previous lemma, f has at most order 1. Theorem 3.1 shows
us that
f (z) = eg(z)
for some entire function g. By the Hadamard Factorization Theorem, g must
be of degree less than or equal to 1. Hence f (z) = aebz for some a, b ∈ C, as we
wanted to show.

4 Conclusion
Hopefully at this point the reader understands how all of this material unifies
itself. Though we had to make a diversion to study the Hadamard and Weier-
strass factorization theorems, they lead us to Buckholtz’s astonishing equiva-
lence result concerning the exponential function. In particular, this shows that
the zeroes of the partial sums of the exponential function increase faster than
those of any other entire function.

12
References
[1] J. D. Buckholtz, A Characterization of the Exponential Series, The American Mathemat-
ical Monthly 73 (Apr., 1966), 121-123.
[2] , Concerning an Approximation of Copson, Proceedings of the American Mathe-
matical Society 14 (Aug., 1963), 564-568.
[3] John B. Conway, Functions of One Complex Variable, Springer-Verlag, New York, 1973.
[4] B. Ja. Levin, Distribution of Zeros of Entire Functions, American Mathematical Society,
United States, 1980.
[5] Peter Walker, The Zeros of the Partial Sums of the Exponential Series, The American
Mathematical Monthly 110 (Apr., 2003), 337-339.

13